An intelligent model for early kick detection based on cost-sensitive learning

Kick detection is crucial for ensuring process safety of drilling operation. Detection of a kick at early stage leaves more time for the drilling crew to take necessary actions. In this work, a novel intelligent model is proposed for early kick detection, which incorporates feature transformation, cost-sensitive dataset construction, and ensemble learning. It applies 7 wellhead feature parameters as input. The model is trained and tested with the field data of a shale gas reservoir in Sichuan. The model performances under different data dimensions and misclassification costs are evaluated. It is found that when the data dimension is 6 and the misclassification cost is 3, the model has the best classification ability (Total Cost=0.9, Accuracy=0.998, Recall=0.990, Precision=0.986). The low false alarm rate helps to minimize wastage of drilling time. The ablation experiment and the comparison with conventional sampling methods unanimously prove the superiority of the proposed model. Datasets with various sizes and imbalance ratios are tested and the model shows satisfactory accuracy. The formula of the optimal misclassification cost is derived for the instruction of field application. The early kick detection performance of the proposed model is better than the existed methods.

Automated summary

Kick detection is crucial for drilling operation safety. In this study, a novel intelligent model is proposed for early kick detection, which incorporates feature transformation, cost-sensitive dataset construction, and ensemble learning. The model shows excellent performance in different data dimensions and misclassification costs, and outperforms conventional methods according to ablation experiments and comparisons. The proposed model demonstrates better early kick detection performance than existing methods.